Most companies can tell you their forecast accuracy last quarter. Very few can tell you what their forecast said in week two of that quarter, which is the only version of the number that could have changed the outcome.
The gap exists because CRMs overwrite. The opportunity record shows today's amount, today's stage, and today's close date. Last month's version is gone unless something captured it. Forecast accuracy tracking is mostly the discipline of capturing what the forecast said before the answer was known.
What has to be captured before tracking is possible?
A weekly snapshot of every open opportunity plus the submitted number at each level of the hierarchy, stored in a table that is never updated in place. Append rows, never overwrite them.Each snapshot row needs the opportunity identifier, amount, stage, close date, owner, forecast category, and the date of the snapshot. Add the submitted forecast for each rep, each manager, and the company, since the difference between a rep's number and their manager's number is one of the more revealing series you can build.
Storing only the roll-up total is the common shortcut and it costs you every diagnostic that matters. A roll-up that moved from $4.1 million to $3.6 million tells you something happened. The opportunity-level snapshot tells you whether deals slipped out, shrank in value, or were removed entirely, and those three have completely different responses.
How do you fix the actuals definition so the history stays comparable?
Pick one closed-won definition, write it down with the date it took effect, and version it rather than editing it. Most accuracy histories are broken by definition drift rather than bad data.The usual sources of drift are a change in what counts as closed, a change in whether multi-year contracts are counted at annual or total value, and a reorganization that moves accounts between segments. Each one silently changes the meaning of the series. Keeping a dated definition file means the trend can be read across the change with a note rather than misread as a real movement.
The bar here is consistency rather than perfection. Everyone believes their CRM data is uniquely bad and that it is the reason they cannot forecast. It is not. Inconsistent data breaks prediction. Consistently imperfect data still produces accurate predictions, because a model can learn a stable error and correct for it.
What views should the tracking system produce?
Four, each answering a different question, and none of them a single-number scorecard.| View | Question it answers | Read it when |
|---|---|---|
| Error by quarter, signed | Are we biased in one direction? | Quarterly, in the retrospective |
| Error by week of quarter | How early does our number become right? | Quarterly, after close |
| Snapshot waterfall | What moved the forecast this week? | Weekly, before the forecast call |
| Segment error trend | Which part of the business is broken? | Quarterly, with cell sizes shown |
How do you read the weekly snapshot waterfall?
Decompose every week-over-week change into created, closed won, closed lost, slipped out, pulled in, and value changed. Six buckets, and they must sum to the total movement or the snapshot has a gap.Slipped out is the bucket that deserves the most attention. When a rep changes a close date, the deal becomes less likely to close even when it stays in commit. A quarter where the forecast holds steady while the slip bucket grows every week is a quarter that is about to break, and the top-line number will not show it until week ten.
The quieter signal is absence of movement. An opportunity with no change in stage, amount, or close date is not stable, it is unattended, and across ORM's customer base the absence of a signal is the earliest slippage signal there is. Track the count of opportunities with no meaningful change in the last 30 days as its own weekly series, since a rising line there tends to lead the slip bucket. Definitions sit in our deal slippage glossary entry.
What does a healthy accuracy trend look like?
Signed error moving toward zero, absolute error shrinking more slowly, and the day-one to day-90 gap closing faster than either. That combination means the process learned to predict rather than to correct.Expect the day-one number to be the hardest to move. Around 90 percent accuracy on new and expansion business is reachable with manual effort, but the manual build goes stale when conditions shift inside the quarter and it takes real work to produce. A model fully trained in 4 to 6 weeks on a company's own historical sales performance updates as the quarter progresses, which is how accuracy holds from day 1 through day 90 rather than arriving in week eleven.
Watch for a pattern where every segment moves the same direction at once. That is rarely a process problem. It usually means something changed in the market or in the business and the forecast is running on old assumptions, whether that is pricing pressure from a new competitor compressing deal sizes or slower buyer decisions stretching cycles. Seasonality does the same thing on a predictable schedule, with Q2 and Q4 typically stronger than Q1 and Q3, and the third month of a quarter stronger than the first two. A tracking system that has not accounted for that will report a bias that is actually a calendar.
Who should own the tracking?
Revenue operations owns the data and the definitions, and the forecast owner presents the trend. Splitting it that way keeps the measurement independent of the person being measured.Put the trend in the same meeting every quarter rather than creating a separate accuracy review, since a standalone meeting gets cancelled by the third quarter. The forecast retrospective and the quarterly business review are both natural homes. More on the surrounding process is in our sales forecasting best practices guide.
Frequently Asked Questions
How often should you snapshot the forecast?
Weekly, on the same day and at the same time, with no exceptions for holidays or quarter ends. An irregular cadence makes week-over-week movement uninterpretable because you cannot tell whether the change came from the pipeline or the calendar.
What has to be stored in a forecast snapshot?
Every open opportunity with amount, stage, close date, owner, and forecast category, plus the submitted number at each level of the hierarchy. Storing only the roll-up number means you can see that a forecast changed but never why.
How many quarters of history do you need before the trend means anything?
You need several quarters before a trend is readable, and a couple of quarters moving the same direction is often still variance, especially with a small number of large deals.
Should accuracy tracking be visible to the sales team?
Yes at the team and segment level, with care at the individual level. Published accuracy trends drive better forecast calls. Individual rankings without context push reps toward sandbagging, which trades one bias for another.
What breaks a forecast accuracy history most often?
Changing a definition mid-stream. Renaming stages, altering what counts as closed, or reorganizing territories all break comparability. Version the definitions with dates so the trend can be read across the change instead of silently distorted by it.
See how ORM turns these insights into action
ORM builds custom revenue forecast models for B2B SaaS companies. Not dashboards. Prescriptive analytics that tell you what to do next.
Schedule a Demo